This Small Business Innovation Research Program (SBIR) Phase I project is aimed at the development and commercialization of a high-performance cloud processing platform for Internet- and mobile-based video services. Target users of this platform would initiate live video streams, or upload video files, through a web portal or application programming interface (API) for processing services (such as transcoding to produce formats and bit-rates suitable for display on mobile devices). Building this platform is challenging due to the need to (a) provide real-time guarantees on shared infrastructure, (b) achieve linear performance scaling, (c) guarantee content security and (d) deliver high availability (99.99%). To address these challenges, the feasibility of (a) a novel, video-aware parallel and distributed stream processing framework (called Scarlet), and (b) a low-overhead isolation framework (called Shield), will be explored in phase I of this project. Scarlet's design and implementation aim for guaranteed real-time processing and linear performance scaling (up to 100 compute units) while delivering professional-grade video quality. Shield should isolate content and resources for real-time media processing on shared infrastructure; content security should be comparable with today's virtual machine (VM) technology but without the associated performance penalties.
The broader impact/commercial potential of this project will be significant as the underlying technology simplifies video processing workflows and reduces associated costs for enterprises that work with large volumes of media content. In particular, the cloud video services resulting from this project will benefit TV broadcasters, Internet-based media companies and mobile video service providers, ultimately benefiting consumers of video in terms of lower prices and better quality. The proposed platform to be developed in this project will be the first to leverage ground-up distributed processing techniques with a high degree of parallelism, scale and reliability to achieve fast turn-around times (>>10x real-time) for media processing. These characteristics will enable this platform to serve as an efficiency enabler for this large and fast-growing Internet video services industry, while also advancing our scientific understanding of cloud technology as applied to media processing. Other emerging vertical segments likely to benefit from this platform include on-line education, surveillance, and augmented reality.
This SBIR project focuses on fundamentally improving technology and reducing costs associated with distributing professional video content to consumers over the Internet. Our goal is the development and commercialization of a high-performance, cloud video processing platform (called "Zipreel Video Cloud" or ZVC) that aids video or media distribution for Internet and mobile clients. Summary: Video content (e.g. a movie) needs to be processed before it can be distributed. This processing is required to convert videos into different resolutions to fit different screen-sizes. It also helps compress videos into different bit-rates (or "quality" levels) to match the different network speeds over which consumers watch this content. From a technical perspective, this processing requires a lot of computing "horsepower", and the challenge is to reduce the time required. From a commercial perspective, reducing the associated costs is also critical when dealing with large media volumes. E.g. consider the problem of processing hundreds of hours of video content in a few hours or even minutes at low cost. Existing solutions solve this problem by treating every video file as an atomic entity and serially processing the files through the fastest available servers. Thus, the processing speed is limited by the fastest processors operating serially on the data. In this project, we explore a "divide-and-conquer" approach that involves breaking each video file into smaller "chunks" that can be processed in parallel. Such an approach allows a cluster of inexpensive small-to-medium-range computers to be used to reduce processing times and costs by an order-of-magnitude. Due to dependencies in input video files, it is non-trivial to break them apart into smaller chunks, process them in parallel, and then put them back together. Solving this problem, and building a reliable and secure software platform around it, are the high-level technical goals of this project. From a commercial perspective, there are three classes of customers who would benefit from the proposed solution: content owners (e.g. Disney and ESPN), content distributors (e.g. Comcast and Verizon) and commercial enterprises that leverage video content for training/education (e.g. Universities and large public/private companies). Content owners and distributors are constantly looking to expand the reach of their assets to as many end platforms as possible. In this context, being able to process their assets to improve that reach and do so rapidly and reliably has a lot of value. Goals and accomplishments from Phase I: In Phase I, as part of exploring feasibility, we focused on "speeding-up" commonly-used processing functions (i.e. transcoding and format conversion) for file-based video content for one set of typical audio/video formats and codecs. Specifically, we wanted to generate professional-grade output videos for Internet/mobile distribution, while supporting rapid turn-around time and high reliability. In addition, we also wanted to provide potential customers using our technology with intuitive user-interfaces. Finally, we targeted incorporating feedback from early adopters via field-trials. We met these goals by accomplishing the following: Designed and implemented a novel, video-aware parallel and distributed software processing framework (called Scarlet) that supports professional-grade output quality, rapid turn-around time and high reliability. Scarlet runs on a grid or cluster of servers. It splits input videos into atomic segments that are processed in parallel before being stitched back together. Scarlet’s differentiating feature is its ability to scale performance linearly with the number of servers in the cluster, which supports order-of-magnitude reduction in turn-around time. Developed user interfaces for Scarlet that ease management and aid automation Engaged with target customers and identified early adopters; a prototype release of Scarlet is being benchmarked at two customer sites; customer feedback will help further fine-tune Scarlet’s features and capabilities Broader/Commercial Impact: The Zipreel product is initially targeted at the broadcast industry which represents a $1.2B annual equipment market that is growing at an average rate of ~8-10% in 2013-2014. The Zipreel product can also enable cost-effective solutions for a variety of government, defense, and educational applications involving large-scale processing of video. Specific project outcomes through the life of the award: An innovative approach to address the basic problem of partitioning video processing in a cluster of computers to achieve order-of-magnitude reduction in turn-around time while not compromising on output video quality. Novel scheduling and load-balancing algorithms that help distribute video processing in a fair manner across the cluster, while minimizing job completion time, and maximizing utilization of cluster resources. A new video-aware parallel and distributed video processing software framework called Scarlet that incorporates the innovations mentioned above to make them available in one software package. Patent filings covering these innovations that advances the state-of-the-art by providing specific and proven guidance on solving this problem. With the SBIR Phase II investment, Zipreel anticipates steady growth in revenues, as well as jobs created, reaching 15M$/yr and ~30 employees in 2018.
